방향각 공간에서의 확산 모델을 이용한 텍스트 기반 2D 인체 포즈 생성

초록

With the recent advances in text-driven generative technologies, research on generating human poses from natural language has attracted increasing attention. This paper proposes a directional-angle-based diffusion model for generating 2D skeleton poses from natural language text. Conventional joint-coordinate-based methods have difficulty explicitly reflecting the geometric constraints of the human body, which may lead to structurally inconsistent poses. To address this limitation, we introduce a directional angle representation that expresses the direction of each bone using azimuth(θ) and elevation(ϕ), and directly model this representation as the output space of a DDPM(Denoising Diffusion Probabilistic Models). The proposed method is based on a two-stage framework that separates text classification from pose generation. A class-conditional diffusion model first generates directional-angle-based poses, and forward kinematics is then applied to reconstruct structurally consistent poses. Experimental results show that the proposed method improves semantic consistency by about 20 points over a coordinate-based baseline and enables structurally stable pose generation.

키워드